Research On Image Phase Recovery Algorithm Based On Non-convex Optimization And Denoising Prior

Phase retrieval refers to the process of recovering the original signal from the amplitude signal obtained in a known sampling system.Due to the insufficient sampling frequency,the existing measurement equipment can only record the amplitude or intensity information of the signal,and the phase information of the signal cannot be obtained directly,and the phase information contains a large amount of structural information,so it is necessary to obtain the phase through the phase retrieval technology.information.Algorithms related to phase retrieval play an irreplaceable role in many imaging systems,and are widely used in engineering and scientific fields,including optics,diffraction imaging,and X-ray crystallography.In this paper,based on the non-convex optimization algorithm,combined with the prior knowledge of denoising,the phase retrieval algorithm will be further studied.The specific research contents are as follows:First,the hard threshold model in the truncated Wirtinger flow algorithm is studied.This method directly uses hard threshold truncation for the truncation boundary of the gradient component.Although the convergence of the algorithm is guaranteed,some gradient information is lost.Aiming at these problems of the truncated Wirtinger flow algorithm,this paper improves the algorithm.The soft threshold model is used to re-weight the gradient components,so that the gradient components of the truncated boundary act on the iterative process in a weighted form.When ensuring the original convergence of the algorithm,Improved image restoration accuracy.Comparative experiments show that the improved algorithm improves the image restoration accuracy while maintaining the convergence speed.Secondly,based on the pr Deep algorithm,a denoising regularization model is studied by using denoising prior knowledge.The denoising regularization model of the pr Deep algorithm uses the denoising operator Dn CNN network to solve the phase retrieval problem,but in dealing with complex noise,the Dn CNN network has to try to design multiple sets of network parameters,which increases the time cost and computational cost.Aiming at this problem of Dn CNN network,this paper improves the original algorithm based on the advantage of FFDNet to deal with complex noise,introduces FFDNet network into denoising regularization model,and reduces the computational cost of the algorithm through the adaptive ability of FFDNet network to complex noise.,to improve the image restoration accuracy.The experimental results show that under various noise intensities,the recovery performance of the improved algorithm is improved,and the iteration efficiency is significantly improved.Finally,based on the BM3D-pr GAMP algorithm,the generalized approximate message passing algorithm and denoising operator BM3 D are studied.Taking advantage of the denoising operator scalability of the GAMP algorithm,the advanced denoising operator is used to replace the BM3 D denoising operator.In this paper,the FFDNet network is introduced into the GAMP algorithm,thereby improving the BM3D-pr GAMP algorithm,and using the advantages of FFDNet in denoising performance to obtain more accurate denoising prior knowledge.In view of the problem that the neural network can only train multiple sets of network parameters for different variance noises,this paper introduces the denoising scaling technology into the FFDNet-pr GAMP algorithm,and uses this technology to improve the adaptability of the algorithm to noise variance.The experimental results show that under different noise variances,the improved algorithm improves the image restoration performance,and the algorithm efficiency also has obvious advantages compared with the original algorithm.